{
  "id": "R2",
  "url": "https://cartforagents.com/research/article-titles-agent-reading-simulation",
  "title": "Which article titles AI agents cite: a simulation across Claude, GPT and Gemini",
  "description": "Plain titles were cited almost every time, clever ones mostly weren't: a simulation of which article titles Claude, GPT and Gemini models cite and choose to read.",
  "answer": "In a supplied-snippet simulation, literal title-and-description packages for three planned Cart for Agents articles were cited in 405 of 432 Claude sessions, 144 of 144 on GPT-6.1-Sol and 143 of 144 on Gemini 3.8 Flash, against 28 of 108, 5 of 36 and 15 of 36 for deliberately clever packages. Asked to pick five of our 84 article titles, the models chose AI-behaviour and research titles for their own reading and practical product-data titles for a Shopify founder: a six-title behaviour group took 27 to 61% of picks for themselves and 0 to 2% for the founder.",
  "take": "Say what the page is. Plain, literal titles were cited almost every time and clever ones mostly weren't, which is why every title on this site reads like a label, not a pun.",
  "summary": "Plain titles that said what an article contained were cited almost every time; clever ones mostly weren't. And when we asked models what to read themselves versus what to send a busy founder, they built two very different reading lists.",
  "claims": [
    {
      "text": "Across three planned articles, the four literal packages (exact contents, question, checklist, exact contents with attribution) were cited in 405 of 432 Claude sessions (three models pooled), 144 of 144 on GPT-6.1-Sol and 143 of 144 on Gemini 3.8 Flash. The clever packages were cited in 28 of 108, 5 of 36 and 15 of 36.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Supplied snippets for planned articles that had not been written or indexed; one question per article and one thin shopper ask. Title and description changed together. Exploratory cells of 12 sessions per model, with no independent confirmation round."
    },
    {
      "text": "For the agentic-commerce question, the clever package was cited in 1 of 36 Claude sessions, 0 of 12 on GPT-6.1-Sol and 4 of 12 on Gemini 3.8 Flash; for the AI-shopping-agent question, 0 of 36, 0 of 12 and 0 of 12. The question and checklist packages were cited in 35 of 36, 12 of 12 and 12 of 12 for the first question.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Small cells: a 12 of 12 result has a Wilson 95% interval of about 76 to 100%. No single literal format won everywhere."
    },
    {
      "text": "For the negative-product-attributes question, the clever package was already cited in 27 of 36 Claude sessions and 11 of 12 on Gemini 3.8 Flash (5 of 12 on GPT-6.1-Sol), and every literal package was cited in every session on every model.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "The real competing search results for that question were a weak topical match, so this is a ceiling effect, not evidence that one headline beat strong competitors."
    },
    {
      "text": "The main outcome counted the two source IDs each model declared. 89 of 540 Claude answers also cited extra sources in the prose (87 of them from Haiku), and 23 of 180 Gemini 3.8 Flash answers had prose citations that did not match the declared IDs; GPT-6.1-Sol had none. Counting any mention in the prose, or only format-compliant answers, left the clever-versus-literal gap in place.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "The format-compliant subsets have unequal model cell sizes and are a sensitivity check, not a new confirmatory result."
    },
    {
      "text": "With all 84 of our article titles shown and five picks per session, a six-title AI-behaviour group fixed before the run took 120 of 450 picks when Claude models chose for their own reading and 3 of 450 when choosing for a Shopify founder (difference 26.0 points, session-bootstrap 95% interval 22.9 to 28.9). On GPT-6.1-Sol it took 72 of 150 against 0 of 150, and on Gemini 3.8 Flash 92 of 150 against 3 of 150.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Role-framed prompts (an assigned \"free time\" role versus helping a described founder). Not human demand, conscious preference or autonomous reading. Five picks in one session are not independent, which is why the interval resamples whole sessions."
    },
    {
      "text": "For the founder, \"Product specifications with units: make numbers comparable\" was picked in 82 of 90 Claude sessions, 23 of 30 on GPT-6.1-Sol and 30 of 30 on Gemini 3.8 Flash, and \"Spec-sold products: build an agent-ready category data model\" in 62 of 90, 30 of 30 and 30 of 30. Neither title was picked once for the models' own reading.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Titles only, no descriptions; titles as they stood on the test date. Picks for a described founder are not that founder's reading or buying behaviour."
    },
    {
      "text": "The most-picked title for Claude's own reading, the product evidence contract (60 of 90), and the retired gap-score audit (57 of 90) were outside the pre-set behaviour group. We did not move them into it after seeing results.",
      "source_title": "Cart for Agents: Which titles agents cite (method)",
      "source_url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Shows the shift is toward research and evaluation themes generally, not only titles about AI behaviour."
    },
    {
      "text": "Every reports that changing one article's title and description to state its contents moved it from 2 of 240 to 240 of 240 citations, and that in an assigned free-time reading role, agents' picks had a correlation of 0.00 with the author's popularity ranking.",
      "source_title": "How to Get AI Agents to Recommend Your Articles (Every)",
      "source_url": "https://every.to/also-true-for-humans/how-to-get-ai-agents-to-recommend-your-articles",
      "source_date": "2026-10-27",
      "method": "controlled-simulation",
      "limitation": "Author-reported supplied-snippet simulation; the date is our check date. The popularity ranking is a proxy, not measured human traffic."
    }
  ],
  "sources": [
    {
      "title": "Cart for Agents: Which titles agents cite (method)",
      "url": "https://cartforagents.com/research/article-titles-agent-reading-simulation#method",
      "accessed": "2026-10-27"
    },
    {
      "title": "How to Get AI Agents to Recommend Your Articles (Every)",
      "url": "https://every.to/also-true-for-humans/how-to-get-ai-agents-to-recommend-your-articles",
      "accessed": "2026-10-27"
    },
    {
      "title": "Cart for Agents: Vague vs specific product facts, simulated picks by Claude, GPT and Gemini models",
      "url": "https://cartforagents.com/research/vague-specific-product-facts-simulation",
      "accessed": "2026-10-27"
    }
  ],
  "limitations": [
    "Supplied-snippet simulation. Models chose among titles and descriptions we gave them; nothing was searched, ranked, opened or read, and the three tested articles had not been written or indexed.",
    "The citation test changed the title and description together, used one question per article and one thin user ask, and ran 12 sessions per model per package. It was exploratory, with no fresh confirmation round.",
    "The reading test assigned roles in the prompt. It measures prompted choices, not human readership, human demand, conscious interests or autonomous agent behaviour.",
    "Five models from three families (three Claude models, GPT-6.1-Sol and Gemini 3.8 Flash) called through APIs at low reasoning effort. Gemini 3.1 Pro was not run on these two tests. API answers do not measure ChatGPT, Gemini apps or Google AI Mode.",
    "The clever packages were deliberately weak test alternatives we wrote, not real competitors' titles or our current titles."
  ],
  "related": [
    "https://cartforagents.com/research/agent-shopping-research",
    "https://cartforagents.com/research/vague-specific-product-facts-simulation",
    "https://cartforagents.com/guides/how-ai-shopping-agents-choose-products",
    "https://cartforagents.com/guides/agent-ready-product-data"
  ],
  "published": "2026-10-27",
  "updated": "2026-10-27",
  "checked": "2026-10-27",
  "kind": "research"
}